[Feature] Avoid some recompiles of ReplayBuffer.extend/sample

This change avoids recompiles for back-to-back calls to
`ReplayBuffer.extend` and `.sample` in cases where `LazyTensorStorage`,
`RoundRobinWriter`, and `RandomSampler` are used and the data type is
either tensor or pytree.

ghstack-source-id: d306cb9
Pull Request resolved: #2504

Author: kurtamohler

test/_utils_internal.py

@@ -5,10 +5,12 @@
 from __future__ import annotations
 
 import contextlib
+import logging
 import os
 
 import os.path
 import time
+import unittest
 from functools import wraps
 
 # Get relative file path
@@ -204,6 +206,31 @@ def f_retry(*args, **kwargs):
     return deco_retry
 
 
+# After calling this function, any log record whose name contains 'record_name'
+# and is emitted from the logger that has qualified name 'logger_qname' is
+# appended to the 'records' list.
+# NOTE: This function is based on testing utilities for 'torch._logging'
+def capture_log_records(records, logger_qname, record_name):
+    assert isinstance(records, list)
+    logger = logging.getLogger(logger_qname)
+
+    class EmitWrapper:
+        def __init__(self, old_emit):
+            self.old_emit = old_emit
+
+        def __call__(self, record):
+            nonlocal records
+            self.old_emit(record)
+            if record_name in record.name:
+                records.append(record)
+
+    for handler in logger.handlers:
+        new_emit = EmitWrapper(handler.emit)
+        contextlib.ExitStack().enter_context(
+            unittest.mock.patch.object(handler, "emit", new_emit)
+        )
+
+
 @pytest.fixture
 def dtype_fixture():
     dtype = torch.get_default_dtype()

test/test_rb.py

@@ -17,7 +17,12 @@
 import pytest
 import torch
 
-from _utils_internal import CARTPOLE_VERSIONED, get_default_devices, make_tc
+from _utils_internal import (
+    capture_log_records,
+    CARTPOLE_VERSIONED,
+    get_default_devices,
+    make_tc,
+)
 
 from mocking_classes import CountingEnv
 from packaging import version
@@ -111,6 +116,7 @@
 _has_gym = importlib.util.find_spec("gym") is not None
 _has_snapshot = importlib.util.find_spec("torchsnapshot") is not None
 _os_is_windows = sys.platform == "win32"
+TORCH_VERSION = version.parse(version.parse(torch.__version__).base_version)
 
 torch_2_3 = version.parse(
     ".".join([str(s) for s in version.parse(str(torch.__version__)).release])
@@ -399,6 +405,75 @@ def data_iter():
         ) if cond else contextlib.nullcontext():
             rb.extend(data2)
 
+    @pytest.mark.skipif(
+        TORCH_VERSION < version.parse("2.5.0"), reason="requires Torch >= 2.5.0"
+    )
+    # Compiling on Windows requires "cl" compiler to be installed.
+    # <https://github.com/pytorch/pytorch/blob/8231180147a096a703d8891756068c89365292e0/torch/_inductor/cpp_builder.py#L143>
+    # Our Windows CI jobs do not have "cl", so skip this test.
+    @pytest.mark.skipif(_os_is_windows, reason="windows tests do not support compile")
+    def test_extend_sample_recompile(
+        self, rb_type, sampler, writer, storage, size, datatype
+    ):
+        if rb_type is not ReplayBuffer:
+            pytest.skip(
+                "Only replay buffer of type 'ReplayBuffer' is currently supported."
+            )
+        if sampler is not RandomSampler:
+            pytest.skip("Only sampler of type 'RandomSampler' is currently supported.")
+        if storage is not LazyTensorStorage:
+            pytest.skip(
+                "Only storage of type 'LazyTensorStorage' is currently supported."
+            )
+        if writer is not RoundRobinWriter:
+            pytest.skip(
+                "Only writer of type 'RoundRobinWriter' is currently supported."
+            )
+        if datatype == "tensordict":
+            pytest.skip("'tensordict' datatype is not currently supported.")
+
+        torch._dynamo.reset_code_caches()
+
+        storage_size = 10 * size
+        rb = self._get_rb(
+            rb_type=rb_type,
+            sampler=sampler,
+            writer=writer,
+            storage=storage,
+            size=storage_size,
+        )
+        data_size = size
+        data = self._get_data(datatype, size=data_size)
+
+        @torch.compile
+        def extend_and_sample(data):
+            rb.extend(data)
+            return rb.sample()
+
+        # Number of times to extend the replay buffer
+        num_extend = 30
+
+        # NOTE: The first two calls to 'extend' and 'sample' currently cause
+        # recompilations, so avoid capturing those for now.
+        num_extend_before_capture = 2
+
+        for _ in range(num_extend_before_capture):
+            extend_and_sample(data)
+
+        try:
+            torch._logging.set_logs(recompiles=True)
+            records = []
+            capture_log_records(records, "torch._dynamo", "recompiles")
+
+            for _ in range(num_extend - num_extend_before_capture):
+                extend_and_sample(data)
+
+            assert len(rb) == storage_size
+            assert len(records) == 0
+
+        finally:
+            torch._logging.set_logs()
+
     def test_sample(self, rb_type, sampler, writer, storage, size, datatype):
         if rb_type is RemoteTensorDictReplayBuffer and _os_is_windows:
             pytest.skip(

test/test_utils.py

@@ -14,11 +14,14 @@
 
 import torch
 
-from _utils_internal import get_default_devices
+from _utils_internal import capture_log_records, get_default_devices
+from packaging import version
 from torchrl._utils import _rng_decorator, get_binary_env_var, implement_for
 
 from torchrl.envs.libs.gym import gym_backend, GymWrapper, set_gym_backend
 
+TORCH_VERSION = version.parse(version.parse(torch.__version__).base_version)
+
 
 @pytest.mark.parametrize("value", ["True", "1", "true"])
 def test_get_binary_env_var_positive(value):
@@ -380,6 +383,34 @@ def test_rng_decorator(device):
         torch.testing.assert_close(s0b, s1b)
 
 
+# Check that 'capture_log_records' captures records emitted when torch
+# recompiles a function.
+@pytest.mark.skipif(
+    TORCH_VERSION < version.parse("2.5.0"), reason="requires Torch >= 2.5.0"
+)
+def test_capture_log_records_recompile():
+    torch.compiler.reset()
+
+    # This function recompiles each time it is called with a different string
+    # input.
+    @torch.compile
+    def str_to_tensor(s):
+        return bytes(s, "utf8")
+
+    str_to_tensor("a")
+
+    try:
+        torch._logging.set_logs(recompiles=True)
+        records = []
+        capture_log_records(records, "torch._dynamo", "recompiles")
+        str_to_tensor("b")
+
+    finally:
+        torch._logging.set_logs()
+
+    assert len(records) == 1
+
+
 if __name__ == "__main__":
     args, unknown = argparse.ArgumentParser().parse_known_args()
     pytest.main([__file__, "--capture", "no", "--exitfirst"] + unknown)

torchrl/data/replay_buffers/storages.py

@@ -144,12 +144,29 @@ def load_state_dict(self, state_dict: Dict[str, Any]) -> None:
     def _empty(self):
         ...
 
+    # NOTE: This property is used to enable compiled Storages. Calling
+    # `len(self)` on a TensorStorage should normally cause a graph break since
+    # it uses a `mp.Value`, and it does cause a break when the `len(self)` call
+    # happens within a method of TensorStorage itself. However, when the
+    # `len(self)` call happens in the Storage base class, for an unknown reason
+    # the compiler doesn't seem to recognize that there should be a graph break,
+    # and the lack of a break causes a recompile each time `len(self)` is called
+    # in this context. Also for an unknown reason, we can force the graph break
+    # to happen if we wrap the `len(self)` call with a `property`-decorated
+    # function. For another unknown reason, if we change
+    # `TensorStorage._len_value` from `mp.Value` to int, it seems like there
+    # should no longer be any need to recompile, but recompiles happen anyway.
+    # Ideally, this should all be investigated and understood in the future.
+    @property
+    def len(self):
+        return len(self)
+
     def _rand_given_ndim(self, batch_size):
         # a method to return random indices given the storage ndim
         if self.ndim == 1:
             return torch.randint(
                 0,
-                len(self),
+                self.len,
                 (batch_size,),
                 generator=self._rng,
                 device=getattr(self, "device", None),